Google Releases Gemini 4 Carbon Topping Benchmarks Restricted to Cyber Defenders

Baoyu (@dotey) · 2026-10-28 18

Google releases Gemini 4 Carbon, ranking #1 on the overall leaderboard. This week, access is restricted to cybersecurity defense teams; general users and developers must wait.

On October 28, Google released its new model, Gemini 4 Carbon, just two weeks after the launch of Gemini 4 Argon. On Artificial Analysis’s third-party comprehensive leaderboard (v5.0), Carbon scored 59 points, taking first place—1 point higher than Claude Opus 5.5 and 2 points higher than GPT-6 Astra.

You won’t be able to use it for now. When Argon was released, Google AI Pro/Ultra subscribers and paid API developers gained immediate access. However, this week, Carbon is available only to "a select group of cybersecurity defenders." It will subsequently be expanded to more partners via the Fairwind program, with developers, enterprises, and general users gaining access later. The official release date for broader availability has not been announced; in his tweet, Sundar Pichai simply stated it would happen "soon."

Fairwind is Google’s initiative to provide early access to new models for cybersecurity defenders, a program that existed during the Argon launch. According to the official website, priority slots are allocated to national cybersecurity agencies, operators of critical infrastructure (such as healthcare, telecommunications, energy, and finance), and large tech platforms. Access requires application through the Google Cloud customer team, background checks, and usage is restricted to internal security teams. General individuals and typical companies cannot participate.

[Pricing]

API pricing remains identical to Argon: introductory rates are $2 per million input tokens and $10 per million output tokens. After the introductory period ends, prices will rise to $4 and $20, respectively. The introductory pricing for Argon expires on November 30, but Google has not specified when the introductory period for Carbon will end. Additionally, it has not been confirmed whether Carbon will be integrated into the Gemini App or which subscription tiers will include it.

[Advantages over Argon and Competitors]

Google states that Carbon can run continuously for days, autonomously changing strategies when encountering dead ends. This makes it suitable for tracing and fixing subtle bugs in large codebases. According to Google’s internal tests, Carbon’s success rate in generating secure code is more than double that of Argon, while hallucinations (unsupported claims) have decreased by approximately half.

The comparison table published by Google covers 17 benchmarks, with Carbon securing first place in 15 of them. It did not win in two categories:

• FrontierSWE 2 (Coding): GPT-6 Astra scored 65.5%, while Carbon scored 61.2%.
• OSWorld (Computer Operation): Claude Opus 5.5 scored 87.2%, while Carbon scored 81.5%.

Note that scores for competitors in the table are largely taken from their own published figures or public leaderboards, though some were independently run by Google.

[High Security Scores: Does This Enable System Hacking?]

The three scores mentioned in Pichai’s tweet correspond to the three steps security teams take to handle vulnerabilities.

1. Finding Vulnerabilities: SEC-bench Pro score of 88.2%. The tasks involve real-world vulnerabilities found in browser JavaScript engines. The model must identify the vulnerability and write a proof-of-concept (PoC) snippet to demonstrate its existence. This benchmark was run by Google; GPT-6 Astra scored 85.4%.
2. Confirming Exploitability: ExploitGym score of 54.8%. In an isolated, air-gapped environment, the model writes code to exploit known vulnerabilities against test programs. Security teams use this step to prioritize fixes, addressing exploitable issues first. GPT-6 Astra scored 42.4%, and Claude Opus 5.5 scored 34.5%.
3. Fixing Vulnerabilities: CWE-bench score of 71.0. Questions and scoring are provided by third-party evaluator Collinear. The model must fix vulnerabilities in real open-source projects. A solution is considered correct only if the attack fails and all original tests pass.

These same capabilities could potentially be used for malicious attacks if they fall into the wrong hands. Google explicitly states that these scores were measured with safety guardrails disabled. Their approach involves initially restricting access to vetted defenders, participating in the US government’s voluntary pre-release model testing, and adjusting safeguards before opening access to developers and general users.

Official Blog: https://t.co/1rdUKZrzxV Evaluation Methodology: https://t.co/uGewSKk8xe

↳ Quote @sundarpichai (Sundar Pichai): I did say we'd be releasing our models periodically...introducing Gemini 4 Carbon, our next frontier model.

Following Argon's release just two weeks ago, Carbon shows the pace at which we’re pushing the frontier, it ranks #1 on the recently updated v5.0 of @ArtificialAnlys Intelligence Index at 59. It also leads across key cybersecurity evaluations scoring 88.2% on SEC-bench Pro for finding vulnerabilities, 54.8% on ExploitGym for exploitation, and 71.0% on CWE-bench for patching.

We’re seeing significant improvements in long-horizon tasks even compared to Argon, and we’ve strengthened frontier safeguards.

As we did with Argon, we’re going to roll it out gradually + responsibly. Some trusted cyber defenders will get access starting this week. We'll be expanding it to more partners through our Fairwind Program, ahead of wider availability soon, so hang tight!